Petascale Simulation Initiative Summary Final Report

The Petascale Simulation Initiative LDRD project began in May 2004 and ended in April 2007. The overall goal of the project was to develop a computational model that would help LLNL simulation codes use petaflop-scale computers effectively, with specific emphasis on multiscale simulation. An early deliverable of the project was a report that surveyed a range of LLNL simulation codes so that the project would better understand their structure and needs. This report was titled 'Survey of Selected LLNL Unclassified Parallel Simulation Codes', UCRL-TR-207192. Based on the findings of this report, the project proceeded along two complementary paths. One goal was to develop a technique to reduce the amount of redundant calculation in a multiscale simulation by caching certain fine-scale simulation results as they were generated and then interpolating new values, where appropriate, based on cached data. This technique, which we call 'Adaptive Sampling', was successfully applied in a materials modeling simulation, and the results are detailed in a journal paper entitled 'Embedded polycrystal plasticity and adaptive sampling', UCRL-JRNL-226715. In this paper we showed an example multiscale simulation in which we were able to achieve a 40x speedup over the standard technique for doing the same problem and with very little loss of accuracy. A second paper that further describes adaptive sampling was recently accepted as a journal publication: 'Adaptive sampling in hierarchical simulation', UCRL-JRNL-232594. The second thrust of the project was to develop a parallel programming model that supports dynamic creation and coordination of sequential and parallel subcomputations to form larger 'federated' simulations. We call this model 'Cooperative Parallelism'. The model is described in 'Overview of the Cooperative Parallel programming model', UCRL-CONF-230029, which was submitted to Supercomputing 2007 (but not accepted for publication). We will be revising the paper and submitting it for publication again. In the mean time the model is under study for use in both AX division and B division, and possibly other programs as well. Finally, a higher-level view of our work appears in book chapter entitled 'Multiscale simulation and petascale computing', UCRL-BOOK-227348, in 'Petascale Computing: Algorithms and Applications'. The description in this chapter sets our work in the larger context of petascale simulation